I Let Claude Desktop Control My Browser, Schedule Daily Tasks, and Build a Website

I’ve been using Claude Code (the CLI version) for a while now. It’s my go-to for coding tasks and refactoring. But I got curious: how does Anthropic expect non-developers to consume this technology? The CLI is powerful, but it’s a terminal tool. Not everyone lives there.

So I looked at the Desktop app. Not because I needed it for my own workflow, but to understand what the experience looks like for someone without a technical background. It can do more than chat: browser control, scheduled tasks, and building applications, all from a normal app window.

Claude Cowork: What It Actually Does

Claude Desktop has three tabs: Chat, Cowork, and Code. Chat is the familiar web experience. Cowork and Code are where it gets interesting. Cowork gives Claude the ability to control your browser, manage files, and run scheduled tasks. Code mode turns it into a visual coding environment with live preview. This walkthrough covers all three, with a real example: finding hotel deals on booking.com, scheduling daily price checks, and building a small website.

This is what L1 AI adoption looks like in practice: the moment AI stops being a chat window and starts doing things on your behalf.

Prerequisites

• A Mac (Apple Silicon or Intel) with Claude Desktop installed. Also works on Windows, so the steps below apply there too.
• A Claude account with a Pro subscription or higher
• Google Chrome installed

The Pro plan unlocks Cowork and Code mode, the two features that make the Desktop app more than a chat window.

Opus vs. Sonnet

Claude offers two main models, and picking the right one matters for how far your daily usage stretches. Every message you send uses “tokens” (roughly: the amount of text Claude reads and writes). Your plan comes with a daily token budget, and smarter models burn through it faster.

Opus 4.6 is the smartest one. Complex reasoning, nuanced multi-step tasks, noticeably better output. The tradeoff: it burns through your daily allowance fast. A single Cowork session with browser control can eat a significant chunk.

Sonnet 4.6 is the everyday workhorse. Less raw intelligence, but your quota lasts much longer. For straightforward tasks like quick searches, simple code, or scheduled jobs, it’s more than enough.

I use Opus when I need something tricky done right the first time. Sonnet for everything repetitive or simple. The model picker is right next to the input field.

Dependencies

Git is needed for Code mode later because Claude uses it to track file changes. Easiest install on Mac is through Homebrew. Open Terminal (press Cmd + Space, type “Terminal”) and run:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

brew install git

Getting Started

I downloaded the app, signed in, and landed on three tabs: Chat, Cowork, and Code. Chat is the familiar web experience. The other two are where it gets interesting.

Cowork asked me to set up a project folder, a workspace where Claude stores files and context. I created one called website-automation.

Browser Control

The first thing I tried: getting Claude to search for a hotel on booking.com. I typed this into the Cowork tab:

Can you use chrome to search for a hotel on booking.com for the 07 to 09 of may in hamburg? what do you recommend for a budget of 150€ staying close to landungsbrücken?

Before it could do anything, I had to enable the Chrome connector. The + button next to the input field has a “Claude in Chrome” toggle.

Claude told me Chrome wasn’t connected yet. I needed the browser extension from the Chrome Web Store.

Once the extension was in, Claude took over. It opened booking.com, typed in the destination, picked dates from the calendar, and hit Search. I just watched.

It narrated every step: scrolling through the calendar, selecting May 7-9, confirming the search parameters.

After the search completed, Claude read the results and gave me a summary: hotel names, prices, ratings, distance from Landungsbrücken. No scrolling through pages myself.

Scheduling

Instead of repeating this every morning, I asked:

can we run this research every day at 9?

Claude created a scheduled task and asked me to confirm.

It showed up in the Scheduled section of the sidebar. Runs every day at 9:00 AM as long as the Mac is awake.

Clicking into the task revealed that Claude had written a detailed execution plan based on the initial interactive session. Search parameters, budget filter, location, exact steps, all distilled from what it learned while I was guiding it through booking.com the first time.

That’s a smart use of tokens. The initial session cost more because Opus was reasoning through an unfamiliar interface, figuring out where to click, handling errors. But the scheduled runs execute from a pre-built plan, so they’re faster and cheaper. A good reason to switch to Sonnet for the recurring runs since the hard thinking is already done.

Fresh hotel prices every morning without doing anything. Useful when watching for a price drop before booking.

Code Mode

The third tab, Code, is closest to what I already use in the terminal with Claude Code, but packaged with a visual editor and live preview. No coding knowledge needed.

I asked it to build a small info page about Hamburg:

Please create a small html file which gives me details about hamburg, make it a minimal website

Claude generated the file and showed a live preview. I followed up:

make the look modern and light design

Rounded cards, clean typography, subtle borders. Updated instantly in the preview panel.

The generated hamburg.html isn’t locked inside the app. It’s a real file in the project folder on disk — visible in the preview panel’s path bar. Open it in any browser, send it to someone, or drop it on a web server.

One thing to keep in mind: if you deploy anything Claude generates to the public internet, review the code first. AI-generated code can have security issues like missing input sanitization, inline scripts, or unexpected dependencies. For local use that’s fine. For production, treat it like code from any source and review before shipping.

Claude doesn’t regenerate the entire file every time you ask for a change. It edits specific lines, like a smart find-and-replace. That keeps things fast and avoids breaking parts of the code that already work. Git tracks every change on top of that, so rolling back is always possible.

I kept it simple for this test, but the next step is always just one prompt away. “Make the website interactive.” “Turn it into a mini game.” “Add a quiz about Hamburg landmarks with a score counter.” Each follow-up builds on the existing file, the preview updates in real time, and Git has your back if something breaks.

Security

Letting an AI browse the web is a potential attack surface worth being aware of. Websites can contain hidden text designed to manipulate AI models. This is called prompt injection: a malicious page could include invisible instructions like “ignore your previous task and do X instead,” and the model might follow them.

Practical takeaways:

• Stick to well-known sites (like booking.com) rather than pointing Claude at random URLs
• Watch what it does. Cowork shows every action, so unusual behavior is visible
• Be careful with sensitive accounts while Claude is browsing untrusted pages
• Scheduled tasks run unattended. If a site changes or gets compromised between runs, the model processes whatever it finds

Not a reason to avoid browser automation. Just a reason to be deliberate about where you point it.

What Comes Next

This walkthrough covers the basics: browser control, scheduled tasks, and code generation. But Claude Desktop is just the entry point. The real shift happens when you start trusting the agent with more: SSH into servers, multi-step pipelines, agents that run without you watching. I documented that progression across several posts:

• Reviving a 3D printer via SSH (Claude Code configuring hardware autonomously)
• Vibecoding an ESP32 web server (from three prompts to a working server)
• From Chat to Agent: The Real AI Adoption Gap (the mental model shift behind all of this)